3D Shape Segmentation And Co-segmentation Via Deep Bayes Inference And Low-rank Representation

3D model segmentation and labeling is a fundamental problem of computer graphics and manufacturing.As consistently labeling a large amount of meshes manually is a tedious procedure accompanied by inevitable mis-labelings,therefore,it is necessary to propose the interpretable automatic segmentation method.However,supervised data-driven methods usually require tedious preprocessing.In addition,in practice,it is difficult to obtain the exact segmentation number and a large number of accurately labeled models,which greatly limits the application of the supervised segmentation methods.By competing with that,the unsupervised 3D model segmentation method is particularly advanced.By transforming the segmentation problem as the clustering problem,we propose two unsupervised clustering algorithms,solving the 3D model semantic segmentation problem.The first method is the kernel low-rank based clustering algorithm.First,we embed the input geometric features into an implicit feature space,then we perform the clustering in the implicit space.In order to cope with the noise of input,we incorporate the block diagonal prior into the prevalent low-rank representation method,obtaining a robust subspace clustering model.With virtue of the block diagonal property and implicit low-rank representation,the proposed model can deal with the noise/outliers and handle input the non-linear geometric feature.Although the first model can deal with the noise well,it cannot be applied to large-scale dataset due to the exist of high dimensional similarity matrix.Moreover,the appropriate implicit feature space is difficult to be defined manually.To relieve these issuers,we propose another method,which makes the implicit features space learnable by a data-driven way,and we assume that the data comes from a Gaussian mixture model in the implicit space.In general,the second method convert the clustering problem into a Bayesian inference problem and we inference the parameters of the learnable distribution by maximizing the log-likelihood.This unsupervised deep Bayesian method is more suitable for largescale datasets.In the experiment,we validate the effectiveness of the proposed methods on 3D model segmentation and image clustering tasks.Compared with the stat-of-art methods,the proposed method achieves better results.